Year: 2026
Type: article
Abstract: Data anonymization in clinical environments requires balancing privacy protection with the preservation of analytical utility. This study compares three hybrid machine learning–based approaches—an autoencoder–GAN model, a CTGAN combined with differential privacy, and a simulated federated-learning scheme with differential privacy—applied to the eICU-CRD clinical dataset. The models were evaluated using structural similarity metrics, correlation preservation, and reidentification risk. Results show that the autoencoder–GAN model provides the highest statistical fidelity, while CTGAN with differential privacy achieves the best trade-off between utility and formal privacy guarantees. The federated approach, although less precise, offers strategic advantages in multi-institutional settings. A functional prototype based on CTGAN + DP was developed and used to generate synthetic tables with quantified deviation from the real data. Findings indicate that generative models with formal privacy mechanisms constitute viable solutions for data anonymization in digital health, provided they are integrated into robust governance and ethical frameworks
Source: Revista Formación Estratégica
Institution: University of La Salle
Language: Spanish
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7221233007